Two-thirds of Earth is under cloud at any moment.1 A satellite sees a ground station for five to ten minutes a pass.2 The decision has to be made onboard.
Hanseatic AI is building KONTOR, a hardware-agnostic inference runtime for the vehicles that cannot phone home. Bench-stage, pre-revenue, and looking for a first integration partner.
KONTOR takes the detection, segmentation and screening models you already trust and runs them inside the power, thermal and radiation budget of the hardware you already fly — rad-tolerant FPGAs, vision processors, shielded commercial modules, embedded GPUs on airframes.
Software only. No box to qualify, no second supply chain, no new mass on the bus.
Quantisation, pruning and distillation tuned against the actual accelerator on the vehicle, to the operating point where accuracy loss stops mattering to the analyst and power starts mattering to the spacecraft.
Cloud screening, change detection and classification at capture time. Empty frames never enter the downlink queue. What reaches the ground is a detection with a chip attached, not a terabyte of ocean.
Bit flips, watchdog resets and thermal throttling are weather, not edge cases. The runtime checkpoints, recovers, and degrades to a known-good state rather than a silent wrong answer.
Event-based imagers report only the pixels that change, at microsecond resolution and milliwatt power3 — the data shape a bandwidth-starved vehicle wants. KONTOR schedules sparse, asynchronous input as a first-class case, not an afterthought to the frame pipeline.